src/gromacs/mdrun/runner.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2011-2019,2020, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 /*! \internal \file
  38  *
  39  * \brief Implements the MD runner routine calling all integrators.
  40  *
  41  * \author David van der Spoel <david.vanderspoel@icm.uu.se>
  42  * \ingroup module_mdrun
  43  */
  44 #include "gmxpre.h"
  45
  46 #include "runner.h"
  47
  48 #include "config.h"
  49
  50 #include <cassert>
  51 #include <cinttypes>
  52 #include <csignal>
  53 #include <cstdlib>
  54 #include <cstring>
  55
  56 #include <algorithm>
  57 #include <memory>
  58
  59 #include "gromacs/commandline/filenm.h"
  60 #include "gromacs/domdec/builder.h"
  61 #include "gromacs/domdec/domdec.h"
  62 #include "gromacs/domdec/domdec_struct.h"
  63 #include "gromacs/domdec/gpuhaloexchange.h"
  64 #include "gromacs/domdec/localatomsetmanager.h"
  65 #include "gromacs/domdec/partition.h"
  66 #include "gromacs/ewald/ewald_utils.h"
  67 #include "gromacs/ewald/pme.h"
  68 #include "gromacs/ewald/pme_gpu_program.h"
  69 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  70 #include "gromacs/fileio/checkpoint.h"
  71 #include "gromacs/fileio/gmxfio.h"
  72 #include "gromacs/fileio/oenv.h"
  73 #include "gromacs/fileio/tpxio.h"
  74 #include "gromacs/gmxlib/network.h"
  75 #include "gromacs/gmxlib/nrnb.h"
  76 #include "gromacs/gpu_utils/gpu_utils.h"
  77 #include "gromacs/hardware/cpuinfo.h"
  78 #include "gromacs/hardware/detecthardware.h"
  79 #include "gromacs/hardware/printhardware.h"
  80 #include "gromacs/imd/imd.h"
  81 #include "gromacs/listed_forces/disre.h"
  82 #include "gromacs/listed_forces/gpubonded.h"
  83 #include "gromacs/listed_forces/orires.h"
  84 #include "gromacs/math/functions.h"
  85 #include "gromacs/math/utilities.h"
  86 #include "gromacs/math/vec.h"
  87 #include "gromacs/mdlib/boxdeformation.h"
  88 #include "gromacs/mdlib/broadcaststructs.h"
  89 #include "gromacs/mdlib/calc_verletbuf.h"
  90 #include "gromacs/mdlib/dispersioncorrection.h"
  91 #include "gromacs/mdlib/enerdata_utils.h"
  92 #include "gromacs/mdlib/force.h"
  93 #include "gromacs/mdlib/forcerec.h"
  94 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  95 #include "gromacs/mdlib/makeconstraints.h"
  96 #include "gromacs/mdlib/md_support.h"
  97 #include "gromacs/mdlib/mdatoms.h"
  98 #include "gromacs/mdlib/membed.h"
  99 #include "gromacs/mdlib/qmmm.h"
 100 #include "gromacs/mdlib/sighandler.h"
 101 #include "gromacs/mdlib/stophandler.h"
 102 #include "gromacs/mdlib/updategroups.h"
 103 #include "gromacs/mdrun/mdmodules.h"
 104 #include "gromacs/mdrun/simulationcontext.h"
 105 #include "gromacs/mdrunutility/handlerestart.h"
 106 #include "gromacs/mdrunutility/logging.h"
 107 #include "gromacs/mdrunutility/multisim.h"
 108 #include "gromacs/mdrunutility/printtime.h"
 109 #include "gromacs/mdrunutility/threadaffinity.h"
 110 #include "gromacs/mdtypes/commrec.h"
 111 #include "gromacs/mdtypes/enerdata.h"
 112 #include "gromacs/mdtypes/fcdata.h"
 113 #include "gromacs/mdtypes/group.h"
 114 #include "gromacs/mdtypes/inputrec.h"
 115 #include "gromacs/mdtypes/md_enums.h"
 116 #include "gromacs/mdtypes/mdrunoptions.h"
 117 #include "gromacs/mdtypes/observableshistory.h"
 118 #include "gromacs/mdtypes/simulation_workload.h"
 119 #include "gromacs/mdtypes/state.h"
 120 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
 121 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 122 #include "gromacs/nbnxm/nbnxm.h"
 123 #include "gromacs/nbnxm/pairlist_tuning.h"
 124 #include "gromacs/pbcutil/pbc.h"
 125 #include "gromacs/pulling/output.h"
 126 #include "gromacs/pulling/pull.h"
 127 #include "gromacs/pulling/pull_rotation.h"
 128 #include "gromacs/restraint/manager.h"
 129 #include "gromacs/restraint/restraintmdmodule.h"
 130 #include "gromacs/restraint/restraintpotential.h"
 131 #include "gromacs/swap/swapcoords.h"
 132 #include "gromacs/taskassignment/decidegpuusage.h"
 133 #include "gromacs/taskassignment/decidesimulationworkload.h"
 134 #include "gromacs/taskassignment/resourcedivision.h"
 135 #include "gromacs/taskassignment/taskassignment.h"
 136 #include "gromacs/taskassignment/usergpuids.h"
 137 #include "gromacs/timing/gpu_timing.h"
 138 #include "gromacs/timing/wallcycle.h"
 139 #include "gromacs/timing/wallcyclereporting.h"
 140 #include "gromacs/topology/mtop_util.h"
 141 #include "gromacs/trajectory/trajectoryframe.h"
 142 #include "gromacs/utility/basenetwork.h"
 143 #include "gromacs/utility/cstringutil.h"
 144 #include "gromacs/utility/exceptions.h"
 145 #include "gromacs/utility/fatalerror.h"
 146 #include "gromacs/utility/filestream.h"
 147 #include "gromacs/utility/gmxassert.h"
 148 #include "gromacs/utility/gmxmpi.h"
 149 #include "gromacs/utility/keyvaluetree.h"
 150 #include "gromacs/utility/logger.h"
 151 #include "gromacs/utility/loggerbuilder.h"
 152 #include "gromacs/utility/mdmodulenotification.h"
 153 #include "gromacs/utility/physicalnodecommunicator.h"
 154 #include "gromacs/utility/pleasecite.h"
 155 #include "gromacs/utility/programcontext.h"
 156 #include "gromacs/utility/smalloc.h"
 157 #include "gromacs/utility/stringutil.h"
 158
 159 #include "isimulator.h"
 160 #include "replicaexchange.h"
 161 #include "simulatorbuilder.h"
 162
 163 #if GMX_FAHCORE
 164 #    include "corewrap.h"
 165 #endif
 166
 167 namespace gmx
 168 {
 169
 170
 171 /*! \brief Manage any development feature flag variables encountered
 172  *
 173  * The use of dev features indicated by environment variables is
 174  * logged in order to ensure that runs with such features enabled can
 175  * be identified from their log and standard output. Any cross
 176  * dependencies are also checked, and if unsatisfied, a fatal error
 177  * issued.
 178  *
 179  * Note that some development features overrides are applied already here:
 180  * the GPU communication flags are set to false in non-tMPI and non-CUDA builds.
 181  *
 182  * \param[in]  mdlog                Logger object.
 183  * \param[in]  useGpuForNonbonded   True if the nonbonded task is offloaded in this run.
 184  * \param[in]  pmeRunMode           The PME run mode for this run
 185  * \returns                         The object populated with development feature flags.
 186  */
 187 static DevelopmentFeatureFlags manageDevelopmentFeatures(const gmx::MDLogger& mdlog,
 188                                                          const bool           useGpuForNonbonded,
 189                                                          const PmeRunMode     pmeRunMode)
 190 {
 191     DevelopmentFeatureFlags devFlags;
 192
 193     // Some builds of GCC 5 give false positive warnings that these
 194     // getenv results are ignored when clearly they are used.
 195 #pragma GCC diagnostic push
 196 #pragma GCC diagnostic ignored "-Wunused-result"
 197     devFlags.enableGpuBufferOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr)
 198                                   && (GMX_GPU == GMX_GPU_CUDA) && useGpuForNonbonded;
 199     devFlags.forceGpuUpdateDefault = (getenv("GMX_FORCE_UPDATE_DEFAULT_GPU") != nullptr);
 200     devFlags.enableGpuHaloExchange =
 201             (getenv("GMX_GPU_DD_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
 202     devFlags.enableGpuPmePPComm =
 203             (getenv("GMX_GPU_PME_PP_COMMS") != nullptr && GMX_THREAD_MPI && (GMX_GPU == GMX_GPU_CUDA));
 204 #pragma GCC diagnostic pop
 205
 206     if (devFlags.enableGpuBufferOps)
 207     {
 208         GMX_LOG(mdlog.warning)
 209                 .asParagraph()
 210                 .appendTextFormatted(
 211                         "This run uses the 'GPU buffer ops' feature, enabled by the "
 212                         "GMX_USE_GPU_BUFFER_OPS environment variable.");
 213     }
 214
 215     if (devFlags.forceGpuUpdateDefault)
 216     {
 217         GMX_LOG(mdlog.warning)
 218                 .asParagraph()
 219                 .appendTextFormatted(
 220                         "This run will default to '-update gpu' as requested by the "
 221                         "GMX_FORCE_UPDATE_DEFAULT_GPU environment variable. GPU update with domain "
 222                         "decomposition lacks substantial testing and should be used with caution.");
 223     }
 224
 225     if (devFlags.enableGpuHaloExchange)
 226     {
 227         if (useGpuForNonbonded)
 228         {
 229             if (!devFlags.enableGpuBufferOps)
 230             {
 231                 GMX_LOG(mdlog.warning)
 232                         .asParagraph()
 233                         .appendTextFormatted(
 234                                 "Enabling GPU buffer operations required by GMX_GPU_DD_COMMS "
 235                                 "(equivalent with GMX_USE_GPU_BUFFER_OPS=1).");
 236                 devFlags.enableGpuBufferOps = true;
 237             }
 238             GMX_LOG(mdlog.warning)
 239                     .asParagraph()
 240                     .appendTextFormatted(
 241                             "This run uses the 'GPU halo exchange' feature, enabled by the "
 242                             "GMX_GPU_DD_COMMS environment variable.");
 243         }
 244         else
 245         {
 246             GMX_LOG(mdlog.warning)
 247                     .asParagraph()
 248                     .appendTextFormatted(
 249                             "GMX_GPU_DD_COMMS environment variable detected, but the 'GPU "
 250                             "halo exchange' feature will not be enabled as nonbonded interactions "
 251                             "are not offloaded.");
 252             devFlags.enableGpuHaloExchange = false;
 253         }
 254     }
 255
 256     if (devFlags.enableGpuPmePPComm)
 257     {
 258         if (pmeRunMode == PmeRunMode::GPU)
 259         {
 260             GMX_LOG(mdlog.warning)
 261                     .asParagraph()
 262                     .appendTextFormatted(
 263                             "This run uses the 'GPU PME-PP communications' feature, enabled "
 264                             "by the GMX_GPU_PME_PP_COMMS environment variable.");
 265         }
 266         else
 267         {
 268             std::string clarification;
 269             if (pmeRunMode == PmeRunMode::Mixed)
 270             {
 271                 clarification =
 272                         "PME FFT and gather are not offloaded to the GPU (PME is running in mixed "
 273                         "mode).";
 274             }
 275             else
 276             {
 277                 clarification = "PME is not offloaded to the GPU.";
 278             }
 279             GMX_LOG(mdlog.warning)
 280                     .asParagraph()
 281                     .appendText(
 282                             "GMX_GPU_PME_PP_COMMS environment variable detected, but the "
 283                             "'GPU PME-PP communications' feature was not enabled as "
 284                             + clarification);
 285             devFlags.enableGpuPmePPComm = false;
 286         }
 287     }
 288
 289     return devFlags;
 290 }
 291
 292 /*! \brief Barrier for safe simultaneous thread access to mdrunner data
 293  *
 294  * Used to ensure that the master thread does not modify mdrunner during copy
 295  * on the spawned threads. */
 296 static void threadMpiMdrunnerAccessBarrier()
 297 {
 298 #if GMX_THREAD_MPI
 299     MPI_Barrier(MPI_COMM_WORLD);
 300 #endif
 301 }
 302
 303 Mdrunner Mdrunner::cloneOnSpawnedThread() const
 304 {
 305     auto newRunner = Mdrunner(std::make_unique<MDModules>());
 306
 307     // All runners in the same process share a restraint manager resource because it is
 308     // part of the interface to the client code, which is associated only with the
 309     // original thread. Handles to the same resources can be obtained by copy.
 310     {
 311         newRunner.restraintManager_ = std::make_unique<RestraintManager>(*restraintManager_);
 312     }
 313
 314     // Copy members of master runner.
 315     // \todo Replace with builder when Simulation context and/or runner phases are better defined.
 316     // Ref https://redmine.gromacs.org/issues/2587 and https://redmine.gromacs.org/issues/2375
 317     newRunner.hw_opt    = hw_opt;
 318     newRunner.filenames = filenames;
 319
 320     newRunner.oenv            = oenv;
 321     newRunner.mdrunOptions    = mdrunOptions;
 322     newRunner.domdecOptions   = domdecOptions;
 323     newRunner.nbpu_opt        = nbpu_opt;
 324     newRunner.pme_opt         = pme_opt;
 325     newRunner.pme_fft_opt     = pme_fft_opt;
 326     newRunner.bonded_opt      = bonded_opt;
 327     newRunner.update_opt      = update_opt;
 328     newRunner.nstlist_cmdline = nstlist_cmdline;
 329     newRunner.replExParams    = replExParams;
 330     newRunner.pforce          = pforce;
 331     // Give the spawned thread the newly created valid communicator
 332     // for the simulation.
 333     newRunner.communicator        = MPI_COMM_WORLD;
 334     newRunner.ms                  = ms;
 335     newRunner.startingBehavior    = startingBehavior;
 336     newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>(*stopHandlerBuilder_);
 337
 338     threadMpiMdrunnerAccessBarrier();
 339
 340     return newRunner;
 341 }
 342
 343 /*! \brief The callback used for running on spawned threads.
 344  *
 345  * Obtains the pointer to the master mdrunner object from the one
 346  * argument permitted to the thread-launch API call, copies it to make
 347  * a new runner for this thread, reinitializes necessary data, and
 348  * proceeds to the simulation. */
 349 static void mdrunner_start_fn(const void* arg)
 350 {
 351     try
 352     {
 353         auto masterMdrunner = reinterpret_cast<const gmx::Mdrunner*>(arg);
 354         /* copy the arg list to make sure that it's thread-local. This
 355            doesn't copy pointed-to items, of course; fnm, cr and fplog
 356            are reset in the call below, all others should be const. */
 357         gmx::Mdrunner mdrunner = masterMdrunner->cloneOnSpawnedThread();
 358         mdrunner.mdrunner();
 359     }
 360     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 361 }
 362
 363
 364 void Mdrunner::spawnThreads(int numThreadsToLaunch)
 365 {
 366 #if GMX_THREAD_MPI
 367     /* now spawn new threads that start mdrunner_start_fn(), while
 368        the main thread returns. Thread affinity is handled later. */
 369     if (tMPI_Init_fn(TRUE, numThreadsToLaunch, TMPI_AFFINITY_NONE, mdrunner_start_fn,
 370                      static_cast<const void*>(this))
 371         != TMPI_SUCCESS)
 372     {
 373         GMX_THROW(gmx::InternalError("Failed to spawn thread-MPI threads"));
 374     }
 375
 376     // Give the master thread the newly created valid communicator for
 377     // the simulation.
 378     communicator = MPI_COMM_WORLD;
 379     threadMpiMdrunnerAccessBarrier();
 380 #else
 381     GMX_UNUSED_VALUE(numThreadsToLaunch);
 382     GMX_UNUSED_VALUE(mdrunner_start_fn);
 383 #endif
 384 }
 385
 386 } // namespace gmx
 387
 388 /*! \brief Initialize variables for Verlet scheme simulation */
 389 static void prepare_verlet_scheme(FILE*               fplog,
 390                                   t_commrec*          cr,
 391                                   t_inputrec*         ir,
 392                                   int                 nstlist_cmdline,
 393                                   const gmx_mtop_t*   mtop,
 394                                   const matrix        box,
 395                                   bool                makeGpuPairList,
 396                                   const gmx::CpuInfo& cpuinfo)
 397 {
 398     /* For NVE simulations, we will retain the initial list buffer */
 399     if (EI_DYNAMICS(ir->eI) && ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
 400     {
 401         /* Update the Verlet buffer size for the current run setup */
 402
 403         /* Here we assume SIMD-enabled kernels are being used. But as currently
 404          * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
 405          * and 4x2 gives a larger buffer than 4x4, this is ok.
 406          */
 407         ListSetupType listType =
 408                 (makeGpuPairList ? ListSetupType::Gpu : ListSetupType::CpuSimdWhenSupported);
 409         VerletbufListSetup listSetup = verletbufGetSafeListSetup(listType);
 410
 411         const real rlist_new =
 412                 calcVerletBufferSize(*mtop, det(box), *ir, ir->nstlist, ir->nstlist - 1, -1, listSetup);
 413
 414         if (rlist_new != ir->rlist)
 415         {
 416             if (fplog != nullptr)
 417             {
 418                 fprintf(fplog,
 419                         "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
 420                         ir->rlist, rlist_new, listSetup.cluster_size_i, listSetup.cluster_size_j);
 421             }
 422             ir->rlist = rlist_new;
 423         }
 424     }
 425
 426     if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
 427     {
 428         gmx_fatal(FARGS, "Can not set nstlist without %s",
 429                   !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
 430     }
 431
 432     if (EI_DYNAMICS(ir->eI))
 433     {
 434         /* Set or try nstlist values */
 435         increaseNstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, makeGpuPairList, cpuinfo);
 436     }
 437 }
 438
 439 /*! \brief Override the nslist value in inputrec
 440  *
 441  * with value passed on the command line (if any)
 442  */
 443 static void override_nsteps_cmdline(const gmx::MDLogger& mdlog, int64_t nsteps_cmdline, t_inputrec* ir)
 444 {
 445     assert(ir);
 446
 447     /* override with anything else than the default -2 */
 448     if (nsteps_cmdline > -2)
 449     {
 450         char sbuf_steps[STEPSTRSIZE];
 451         char sbuf_msg[STRLEN];
 452
 453         ir->nsteps = nsteps_cmdline;
 454         if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
 455         {
 456             sprintf(sbuf_msg,
 457                     "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
 458                     gmx_step_str(nsteps_cmdline, sbuf_steps), fabs(nsteps_cmdline * ir->delta_t));
 459         }
 460         else
 461         {
 462             sprintf(sbuf_msg, "Overriding nsteps with value passed on the command line: %s steps",
 463                     gmx_step_str(nsteps_cmdline, sbuf_steps));
 464         }
 465
 466         GMX_LOG(mdlog.warning).asParagraph().appendText(sbuf_msg);
 467     }
 468     else if (nsteps_cmdline < -2)
 469     {
 470         gmx_fatal(FARGS, "Invalid nsteps value passed on the command line: %" PRId64, nsteps_cmdline);
 471     }
 472     /* Do nothing if nsteps_cmdline == -2 */
 473 }
 474
 475 namespace gmx
 476 {
 477
 478 /*! \brief Return whether GPU acceleration of nonbondeds is supported with the given settings.
 479  *
 480  * If not, and if a warning may be issued, logs a warning about
 481  * falling back to CPU code. With thread-MPI, only the first
 482  * call to this function should have \c issueWarning true. */
 483 static bool gpuAccelerationOfNonbondedIsUseful(const MDLogger& mdlog, const t_inputrec& ir, bool issueWarning)
 484 {
 485     bool        gpuIsUseful = true;
 486     std::string warning;
 487
 488     if (ir.opts.ngener - ir.nwall > 1)
 489     {
 490         /* The GPU code does not support more than one energy group.
 491          * If the user requested GPUs explicitly, a fatal error is given later.
 492          */
 493         gpuIsUseful = false;
 494         warning =
 495                 "Multiple energy groups is not implemented for GPUs, falling back to the CPU. "
 496                 "For better performance, run on the GPU without energy groups and then do "
 497                 "gmx mdrun -rerun option on the trajectory with an energy group .tpr file.";
 498     }
 499
 500     if (EI_TPI(ir.eI))
 501     {
 502         gpuIsUseful = false;
 503         warning     = "TPI is not implemented for GPUs.";
 504     }
 505
 506     if (!gpuIsUseful && issueWarning)
 507     {
 508         GMX_LOG(mdlog.warning).asParagraph().appendText(warning);
 509     }
 510
 511     return gpuIsUseful;
 512 }
 513
 514 //! Initializes the logger for mdrun.
 515 static gmx::LoggerOwner buildLogger(FILE* fplog, const bool isSimulationMasterRank)
 516 {
 517     gmx::LoggerBuilder builder;
 518     if (fplog != nullptr)
 519     {
 520         builder.addTargetFile(gmx::MDLogger::LogLevel::Info, fplog);
 521     }
 522     if (isSimulationMasterRank)
 523     {
 524         builder.addTargetStream(gmx::MDLogger::LogLevel::Warning, &gmx::TextOutputFile::standardError());
 525     }
 526     return builder.build();
 527 }
 528
 529 //! Make a TaskTarget from an mdrun argument string.
 530 static TaskTarget findTaskTarget(const char* optionString)
 531 {
 532     TaskTarget returnValue = TaskTarget::Auto;
 533
 534     if (strncmp(optionString, "auto", 3) == 0)
 535     {
 536         returnValue = TaskTarget::Auto;
 537     }
 538     else if (strncmp(optionString, "cpu", 3) == 0)
 539     {
 540         returnValue = TaskTarget::Cpu;
 541     }
 542     else if (strncmp(optionString, "gpu", 3) == 0)
 543     {
 544         returnValue = TaskTarget::Gpu;
 545     }
 546     else
 547     {
 548         GMX_ASSERT(false, "Option string should have been checked for sanity already");
 549     }
 550
 551     return returnValue;
 552 }
 553
 554 //! Finish run, aggregate data to print performance info.
 555 static void finish_run(FILE*                     fplog,
 556                        const gmx::MDLogger&      mdlog,
 557                        const t_commrec*          cr,
 558                        const t_inputrec*         inputrec,
 559                        t_nrnb                    nrnb[],
 560                        gmx_wallcycle_t           wcycle,
 561                        gmx_walltime_accounting_t walltime_accounting,
 562                        nonbonded_verlet_t*       nbv,
 563                        const gmx_pme_t*          pme,
 564                        gmx_bool                  bWriteStat)
 565 {
 566     double delta_t = 0;
 567     double nbfs = 0, mflop = 0;
 568     double elapsed_time, elapsed_time_over_all_ranks, elapsed_time_over_all_threads,
 569             elapsed_time_over_all_threads_over_all_ranks;
 570     /* Control whether it is valid to print a report. Only the
 571        simulation master may print, but it should not do so if the run
 572        terminated e.g. before a scheduled reset step. This is
 573        complicated by the fact that PME ranks are unaware of the
 574        reason why they were sent a pmerecvqxFINISH. To avoid
 575        communication deadlocks, we always do the communication for the
 576        report, even if we've decided not to write the report, because
 577        how long it takes to finish the run is not important when we've
 578        decided not to report on the simulation performance.
 579
 580        Further, we only report performance for dynamical integrators,
 581        because those are the only ones for which we plan to
 582        consider doing any optimizations. */
 583     bool printReport = EI_DYNAMICS(inputrec->eI) && SIMMASTER(cr);
 584
 585     if (printReport && !walltime_accounting_get_valid_finish(walltime_accounting))
 586     {
 587         GMX_LOG(mdlog.warning)
 588                 .asParagraph()
 589                 .appendText("Simulation ended prematurely, no performance report will be written.");
 590         printReport = false;
 591     }
 592
 593     t_nrnb*                 nrnb_tot;
 594     std::unique_ptr<t_nrnb> nrnbTotalStorage;
 595     if (cr->nnodes > 1)
 596     {
 597         nrnbTotalStorage = std::make_unique<t_nrnb>();
 598         nrnb_tot         = nrnbTotalStorage.get();
 599 #if GMX_MPI
 600         MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
 601 #endif
 602     }
 603     else
 604     {
 605         nrnb_tot = nrnb;
 606     }
 607
 608     elapsed_time = walltime_accounting_get_time_since_reset(walltime_accounting);
 609     elapsed_time_over_all_threads =
 610             walltime_accounting_get_time_since_reset_over_all_threads(walltime_accounting);
 611     if (cr->nnodes > 1)
 612     {
 613 #if GMX_MPI
 614         /* reduce elapsed_time over all MPI ranks in the current simulation */
 615         MPI_Allreduce(&elapsed_time, &elapsed_time_over_all_ranks, 1, MPI_DOUBLE, MPI_SUM,
 616                       cr->mpi_comm_mysim);
 617         elapsed_time_over_all_ranks /= cr->nnodes;
 618         /* Reduce elapsed_time_over_all_threads over all MPI ranks in the
 619          * current simulation. */
 620         MPI_Allreduce(&elapsed_time_over_all_threads, &elapsed_time_over_all_threads_over_all_ranks,
 621                       1, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mysim);
 622 #endif
 623     }
 624     else
 625     {
 626         elapsed_time_over_all_ranks                  = elapsed_time;
 627         elapsed_time_over_all_threads_over_all_ranks = elapsed_time_over_all_threads;
 628     }
 629
 630     if (printReport)
 631     {
 632         print_flop(fplog, nrnb_tot, &nbfs, &mflop);
 633     }
 634
 635     if (thisRankHasDuty(cr, DUTY_PP) && DOMAINDECOMP(cr))
 636     {
 637         print_dd_statistics(cr, inputrec, fplog);
 638     }
 639
 640     /* TODO Move the responsibility for any scaling by thread counts
 641      * to the code that handled the thread region, so that there's a
 642      * mechanism to keep cycle counting working during the transition
 643      * to task parallelism. */
 644     int nthreads_pp  = gmx_omp_nthreads_get(emntNonbonded);
 645     int nthreads_pme = gmx_omp_nthreads_get(emntPME);
 646     wallcycle_scale_by_num_threads(wcycle, thisRankHasDuty(cr, DUTY_PME) && !thisRankHasDuty(cr, DUTY_PP),
 647                                    nthreads_pp, nthreads_pme);
 648     auto cycle_sum(wallcycle_sum(cr, wcycle));
 649
 650     if (printReport)
 651     {
 652         auto nbnxn_gpu_timings =
 653                 (nbv != nullptr && nbv->useGpu()) ? Nbnxm::gpu_get_timings(nbv->gpu_nbv) : nullptr;
 654         gmx_wallclock_gpu_pme_t pme_gpu_timings = {};
 655
 656         if (pme_gpu_task_enabled(pme))
 657         {
 658             pme_gpu_get_timings(pme, &pme_gpu_timings);
 659         }
 660         wallcycle_print(fplog, mdlog, cr->nnodes, cr->npmenodes, nthreads_pp, nthreads_pme,
 661                         elapsed_time_over_all_ranks, wcycle, cycle_sum, nbnxn_gpu_timings,
 662                         &pme_gpu_timings);
 663
 664         if (EI_DYNAMICS(inputrec->eI))
 665         {
 666             delta_t = inputrec->delta_t;
 667         }
 668
 669         if (fplog)
 670         {
 671             print_perf(fplog, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
 672                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
 673                        delta_t, nbfs, mflop);
 674         }
 675         if (bWriteStat)
 676         {
 677             print_perf(stderr, elapsed_time_over_all_threads_over_all_ranks, elapsed_time_over_all_ranks,
 678                        walltime_accounting_get_nsteps_done_since_reset(walltime_accounting),
 679                        delta_t, nbfs, mflop);
 680         }
 681     }
 682 }
 683
 684 int Mdrunner::mdrunner()
 685 {
 686     matrix                    box;
 687     t_forcerec*               fr               = nullptr;
 688     t_fcdata*                 fcd              = nullptr;
 689     real                      ewaldcoeff_q     = 0;
 690     real                      ewaldcoeff_lj    = 0;
 691     int                       nChargePerturbed = -1, nTypePerturbed = 0;
 692     gmx_wallcycle_t           wcycle;
 693     gmx_walltime_accounting_t walltime_accounting = nullptr;
 694     gmx_membed_t*             membed              = nullptr;
 695     gmx_hw_info_t*            hwinfo              = nullptr;
 696
 697     /* CAUTION: threads may be started later on in this function, so
 698        cr doesn't reflect the final parallel state right now */
 699     gmx_mtop_t mtop;
 700
 701     /* TODO: inputrec should tell us whether we use an algorithm, not a file option */
 702     const bool doEssentialDynamics = opt2bSet("-ei", filenames.size(), filenames.data());
 703     const bool doMembed            = opt2bSet("-membed", filenames.size(), filenames.data());
 704     const bool doRerun             = mdrunOptions.rerun;
 705
 706     // Handle task-assignment related user options.
 707     EmulateGpuNonbonded emulateGpuNonbonded =
 708             (getenv("GMX_EMULATE_GPU") != nullptr ? EmulateGpuNonbonded::Yes : EmulateGpuNonbonded::No);
 709
 710     std::vector<int> userGpuTaskAssignment;
 711     try
 712     {
 713         userGpuTaskAssignment = parseUserTaskAssignmentString(hw_opt.userGpuTaskAssignment);
 714     }
 715     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 716     auto nonbondedTarget = findTaskTarget(nbpu_opt);
 717     auto pmeTarget       = findTaskTarget(pme_opt);
 718     auto pmeFftTarget    = findTaskTarget(pme_fft_opt);
 719     auto bondedTarget    = findTaskTarget(bonded_opt);
 720     auto updateTarget    = findTaskTarget(update_opt);
 721
 722     FILE* fplog = nullptr;
 723     // If we are appending, we don't write log output because we need
 724     // to check that the old log file matches what the checkpoint file
 725     // expects. Otherwise, we should start to write log output now if
 726     // there is a file ready for it.
 727     if (logFileHandle != nullptr && startingBehavior != StartingBehavior::RestartWithAppending)
 728     {
 729         fplog = gmx_fio_getfp(logFileHandle);
 730     }
 731     const bool       isSimulationMasterRank = findIsSimulationMasterRank(ms, communicator);
 732     gmx::LoggerOwner logOwner(buildLogger(fplog, isSimulationMasterRank));
 733     gmx::MDLogger    mdlog(logOwner.logger());
 734
 735     // TODO The thread-MPI master rank makes a working
 736     // PhysicalNodeCommunicator here, but it gets rebuilt by all ranks
 737     // after the threads have been launched. This works because no use
 738     // is made of that communicator until after the execution paths
 739     // have rejoined. But it is likely that we can improve the way
 740     // this is expressed, e.g. by expressly running detection only the
 741     // master rank for thread-MPI, rather than relying on the mutex
 742     // and reference count.
 743     PhysicalNodeCommunicator physicalNodeComm(communicator, gmx_physicalnode_id_hash());
 744     hwinfo = gmx_detect_hardware(mdlog, physicalNodeComm);
 745
 746     gmx_print_detected_hardware(fplog, isSimulationMasterRank && isMasterSim(ms), mdlog, hwinfo);
 747
 748     std::vector<int> gpuIdsToUse = makeGpuIdsToUse(hwinfo->gpu_info, hw_opt.gpuIdsAvailable);
 749
 750     // Print citation requests after all software/hardware printing
 751     pleaseCiteGromacs(fplog);
 752
 753     // TODO Replace this by unique_ptr once t_inputrec is C++
 754     t_inputrec               inputrecInstance;
 755     t_inputrec*              inputrec = nullptr;
 756     std::unique_ptr<t_state> globalState;
 757
 758     auto partialDeserializedTpr = std::make_unique<PartialDeserializedTprFile>();
 759
 760     if (isSimulationMasterRank)
 761     {
 762         /* Only the master rank has the global state */
 763         globalState = std::make_unique<t_state>();
 764
 765         /* Read (nearly) all data required for the simulation
 766          * and keep the partly serialized tpr contents to send to other ranks later
 767          */
 768         *partialDeserializedTpr = read_tpx_state(ftp2fn(efTPR, filenames.size(), filenames.data()),
 769                                                  &inputrecInstance, globalState.get(), &mtop);
 770         inputrec                = &inputrecInstance;
 771     }
 772
 773     /* Check and update the hardware options for internal consistency */
 774     checkAndUpdateHardwareOptions(mdlog, &hw_opt, isSimulationMasterRank, domdecOptions.numPmeRanks,
 775                                   inputrec);
 776
 777     if (GMX_THREAD_MPI && isSimulationMasterRank)
 778     {
 779         bool useGpuForNonbonded = false;
 780         bool useGpuForPme       = false;
 781         try
 782         {
 783             GMX_RELEASE_ASSERT(inputrec != nullptr, "Keep the compiler happy");
 784
 785             // If the user specified the number of ranks, then we must
 786             // respect that, but in default mode, we need to allow for
 787             // the number of GPUs to choose the number of ranks.
 788             auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
 789             useGpuForNonbonded         = decideWhetherToUseGpusForNonbondedWithThreadMpi(
 790                     nonbondedTarget, gpuIdsToUse, userGpuTaskAssignment, emulateGpuNonbonded,
 791                     canUseGpuForNonbonded,
 792                     gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, GMX_THREAD_MPI),
 793                     hw_opt.nthreads_tmpi);
 794             useGpuForPme = decideWhetherToUseGpusForPmeWithThreadMpi(
 795                     useGpuForNonbonded, pmeTarget, gpuIdsToUse, userGpuTaskAssignment, *hwinfo,
 796                     *inputrec, mtop, hw_opt.nthreads_tmpi, domdecOptions.numPmeRanks);
 797         }
 798         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 799
 800         /* Determine how many thread-MPI ranks to start.
 801          *
 802          * TODO Over-writing the user-supplied value here does
 803          * prevent any possible subsequent checks from working
 804          * correctly. */
 805         hw_opt.nthreads_tmpi = get_nthreads_mpi(hwinfo, &hw_opt, gpuIdsToUse, useGpuForNonbonded,
 806                                                 useGpuForPme, inputrec, &mtop, mdlog, doMembed);
 807
 808         // Now start the threads for thread MPI.
 809         spawnThreads(hw_opt.nthreads_tmpi);
 810         // The spawned threads enter mdrunner() and execution of
 811         // master and spawned threads joins at the end of this block.
 812         physicalNodeComm = PhysicalNodeCommunicator(communicator, gmx_physicalnode_id_hash());
 813     }
 814
 815     GMX_RELEASE_ASSERT(communicator == MPI_COMM_WORLD, "Must have valid world communicator");
 816     CommrecHandle crHandle = init_commrec(communicator, ms);
 817     t_commrec*    cr       = crHandle.get();
 818     GMX_RELEASE_ASSERT(cr != nullptr, "Must have valid commrec");
 819
 820     if (PAR(cr))
 821     {
 822         /* now broadcast everything to the non-master nodes/threads: */
 823         if (!isSimulationMasterRank)
 824         {
 825             inputrec = &inputrecInstance;
 826         }
 827         init_parallel(cr, inputrec, &mtop, partialDeserializedTpr.get());
 828     }
 829     GMX_RELEASE_ASSERT(inputrec != nullptr, "All ranks should have a valid inputrec now");
 830     partialDeserializedTpr.reset(nullptr);
 831
 832     // Now the number of ranks is known to all ranks, and each knows
 833     // the inputrec read by the master rank. The ranks can now all run
 834     // the task-deciding functions and will agree on the result
 835     // without needing to communicate.
 836     //
 837     // TODO Should we do the communication in debug mode to support
 838     // having an assertion?
 839     const bool useDomainDecomposition = (PAR(cr) && !(EI_TPI(inputrec->eI) || inputrec->eI == eiNM));
 840
 841     // Note that these variables describe only their own node.
 842     //
 843     // Note that when bonded interactions run on a GPU they always run
 844     // alongside a nonbonded task, so do not influence task assignment
 845     // even though they affect the force calculation workload.
 846     bool useGpuForNonbonded = false;
 847     bool useGpuForPme       = false;
 848     bool useGpuForBonded    = false;
 849     bool useGpuForUpdate    = false;
 850     bool gpusWereDetected   = hwinfo->ngpu_compatible_tot > 0;
 851     try
 852     {
 853         // It's possible that there are different numbers of GPUs on
 854         // different nodes, which is the user's responsibility to
 855         // handle. If unsuitable, we will notice that during task
 856         // assignment.
 857         auto canUseGpuForNonbonded = buildSupportsNonbondedOnGpu(nullptr);
 858         useGpuForNonbonded         = decideWhetherToUseGpusForNonbonded(
 859                 nonbondedTarget, userGpuTaskAssignment, emulateGpuNonbonded, canUseGpuForNonbonded,
 860                 gpuAccelerationOfNonbondedIsUseful(mdlog, *inputrec, !GMX_THREAD_MPI), gpusWereDetected);
 861         useGpuForPme = decideWhetherToUseGpusForPme(
 862                 useGpuForNonbonded, pmeTarget, userGpuTaskAssignment, *hwinfo, *inputrec, mtop,
 863                 cr->nnodes, domdecOptions.numPmeRanks, gpusWereDetected);
 864         auto canUseGpuForBonded = buildSupportsGpuBondeds(nullptr)
 865                                   && inputSupportsGpuBondeds(*inputrec, mtop, nullptr);
 866         useGpuForBonded = decideWhetherToUseGpusForBonded(
 867                 useGpuForNonbonded, useGpuForPme, bondedTarget, canUseGpuForBonded,
 868                 EVDW_PME(inputrec->vdwtype), EEL_PME_EWALD(inputrec->coulombtype),
 869                 domdecOptions.numPmeRanks, gpusWereDetected);
 870     }
 871     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
 872
 873     const PmeRunMode pmeRunMode = determinePmeRunMode(useGpuForPme, pmeFftTarget, *inputrec);
 874
 875     // Initialize development feature flags that enabled by environment variable
 876     // and report those features that are enabled.
 877     const DevelopmentFeatureFlags devFlags =
 878             manageDevelopmentFeatures(mdlog, useGpuForNonbonded, pmeRunMode);
 879
 880     const bool inputIsCompatibleWithModularSimulator = ModularSimulator::isInputCompatible(
 881             false, inputrec, doRerun, mtop, ms, replExParams, nullptr, doEssentialDynamics, doMembed);
 882     const bool useModularSimulator = inputIsCompatibleWithModularSimulator
 883                                      && !(getenv("GMX_DISABLE_MODULAR_SIMULATOR") != nullptr);
 884
 885     // Build restraints.
 886     // TODO: hide restraint implementation details from Mdrunner.
 887     // There is nothing unique about restraints at this point as far as the
 888     // Mdrunner is concerned. The Mdrunner should just be getting a sequence of
 889     // factory functions from the SimulationContext on which to call mdModules_->add().
 890     // TODO: capture all restraints into a single RestraintModule, passed to the runner builder.
 891     for (auto&& restraint : restraintManager_->getRestraints())
 892     {
 893         auto module = RestraintMDModule::create(restraint, restraint->sites());
 894         mdModules_->add(std::move(module));
 895     }
 896
 897     // TODO: Error handling
 898     mdModules_->assignOptionsToModules(*inputrec->params, nullptr);
 899     const auto& mdModulesNotifier = mdModules_->notifier().notifier_;
 900
 901     if (inputrec->internalParameters != nullptr)
 902     {
 903         mdModulesNotifier.notify(*inputrec->internalParameters);
 904     }
 905
 906     if (fplog != nullptr)
 907     {
 908         pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
 909         fprintf(fplog, "\n");
 910     }
 911
 912     if (SIMMASTER(cr))
 913     {
 914         /* In rerun, set velocities to zero if present */
 915         if (doRerun && ((globalState->flags & (1 << estV)) != 0))
 916         {
 917             // rerun does not use velocities
 918             GMX_LOG(mdlog.info)
 919                     .asParagraph()
 920                     .appendText(
 921                             "Rerun trajectory contains velocities. Rerun does only evaluate "
 922                             "potential energy and forces. The velocities will be ignored.");
 923             for (int i = 0; i < globalState->natoms; i++)
 924             {
 925                 clear_rvec(globalState->v[i]);
 926             }
 927             globalState->flags &= ~(1 << estV);
 928         }
 929
 930         /* now make sure the state is initialized and propagated */
 931         set_state_entries(globalState.get(), inputrec, useModularSimulator);
 932     }
 933
 934     /* NM and TPI parallelize over force/energy calculations, not atoms,
 935      * so we need to initialize and broadcast the global state.
 936      */
 937     if (inputrec->eI == eiNM || inputrec->eI == eiTPI)
 938     {
 939         if (!MASTER(cr))
 940         {
 941             globalState = std::make_unique<t_state>();
 942         }
 943         broadcastStateWithoutDynamics(cr, globalState.get());
 944     }
 945
 946     /* A parallel command line option consistency check that we can
 947        only do after any threads have started. */
 948     if (!PAR(cr)
 949         && (domdecOptions.numCells[XX] > 1 || domdecOptions.numCells[YY] > 1
 950             || domdecOptions.numCells[ZZ] > 1 || domdecOptions.numPmeRanks > 0))
 951     {
 952         gmx_fatal(FARGS,
 953                   "The -dd or -npme option request a parallel simulation, "
 954 #if !GMX_MPI
 955                   "but %s was compiled without threads or MPI enabled",
 956                   output_env_get_program_display_name(oenv));
 957 #elif GMX_THREAD_MPI
 958                   "but the number of MPI-threads (option -ntmpi) is not set or is 1");
 959 #else
 960                   "but %s was not started through mpirun/mpiexec or only one rank was requested "
 961                   "through mpirun/mpiexec",
 962                   output_env_get_program_display_name(oenv));
 963 #endif
 964     }
 965
 966     if (doRerun && (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
 967     {
 968         gmx_fatal(FARGS,
 969                   "The .mdp file specified an energy mininization or normal mode algorithm, and "
 970                   "these are not compatible with mdrun -rerun");
 971     }
 972
 973     if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
 974     {
 975         if (domdecOptions.numPmeRanks > 0)
 976         {
 977             gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
 978                                  "PME-only ranks are requested, but the system does not use PME "
 979                                  "for electrostatics or LJ");
 980         }
 981
 982         domdecOptions.numPmeRanks = 0;
 983     }
 984
 985     if (useGpuForNonbonded && domdecOptions.numPmeRanks < 0)
 986     {
 987         /* With NB GPUs we don't automatically use PME-only CPU ranks. PME ranks can
 988          * improve performance with many threads per GPU, since our OpenMP
 989          * scaling is bad, but it's difficult to automate the setup.
 990          */
 991         domdecOptions.numPmeRanks = 0;
 992     }
 993     if (useGpuForPme)
 994     {
 995         if (domdecOptions.numPmeRanks < 0)
 996         {
 997             domdecOptions.numPmeRanks = 0;
 998             // TODO possibly print a note that one can opt-in for a separate PME GPU rank?
 999         }
1000         else
1001         {
1002             GMX_RELEASE_ASSERT(domdecOptions.numPmeRanks <= 1,
1003                                "PME GPU decomposition is not supported");
1004         }
1005     }
1006
1007 #if GMX_FAHCORE
1008     if (MASTER(cr))
1009     {
1010         fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1011     }
1012 #endif
1013
1014     /* NMR restraints must be initialized before load_checkpoint,
1015      * since with time averaging the history is added to t_state.
1016      * For proper consistency check we therefore need to extend
1017      * t_state here.
1018      * So the PME-only nodes (if present) will also initialize
1019      * the distance restraints.
1020      */
1021     snew(fcd, 1);
1022
1023     /* This needs to be called before read_checkpoint to extend the state */
1024     init_disres(fplog, &mtop, inputrec, cr, ms, fcd, globalState.get(), replExParams.exchangeInterval > 0);
1025
1026     init_orires(fplog, &mtop, inputrec, cr, ms, globalState.get(), &(fcd->orires));
1027
1028     auto deform = prepareBoxDeformation(globalState->box, cr, *inputrec);
1029
1030     ObservablesHistory observablesHistory = {};
1031
1032     if (startingBehavior != StartingBehavior::NewSimulation)
1033     {
1034         /* Check if checkpoint file exists before doing continuation.
1035          * This way we can use identical input options for the first and subsequent runs...
1036          */
1037         if (mdrunOptions.numStepsCommandline > -2)
1038         {
1039             /* Temporarily set the number of steps to unlimited to avoid
1040              * triggering the nsteps check in load_checkpoint().
1041              * This hack will go away soon when the -nsteps option is removed.
1042              */
1043             inputrec->nsteps = -1;
1044         }
1045
1046         load_checkpoint(opt2fn_master("-cpi", filenames.size(), filenames.data(), cr),
1047                         logFileHandle, cr, domdecOptions.numCells, inputrec, globalState.get(),
1048                         &observablesHistory, mdrunOptions.reproducible, mdModules_->notifier());
1049
1050         if (startingBehavior == StartingBehavior::RestartWithAppending && logFileHandle)
1051         {
1052             // Now we can start normal logging to the truncated log file.
1053             fplog = gmx_fio_getfp(logFileHandle);
1054             prepareLogAppending(fplog);
1055             logOwner = buildLogger(fplog, MASTER(cr));
1056             mdlog    = logOwner.logger();
1057         }
1058     }
1059
1060     if (mdrunOptions.numStepsCommandline > -2)
1061     {
1062         GMX_LOG(mdlog.info)
1063                 .asParagraph()
1064                 .appendText(
1065                         "The -nsteps functionality is deprecated, and may be removed in a future "
1066                         "version. "
1067                         "Consider using gmx convert-tpr -nsteps or changing the appropriate .mdp "
1068                         "file field.");
1069     }
1070     /* override nsteps with value set on the commandline */
1071     override_nsteps_cmdline(mdlog, mdrunOptions.numStepsCommandline, inputrec);
1072
1073     if (SIMMASTER(cr))
1074     {
1075         copy_mat(globalState->box, box);
1076     }
1077
1078     if (PAR(cr))
1079     {
1080         gmx_bcast(sizeof(box), box, cr);
1081     }
1082
1083     if (inputrec->cutoff_scheme != ecutsVERLET)
1084     {
1085         gmx_fatal(FARGS,
1086                   "This group-scheme .tpr file can no longer be run by mdrun. Please update to the "
1087                   "Verlet scheme, or use an earlier version of GROMACS if necessary.");
1088     }
1089     /* Update rlist and nstlist. */
1090     prepare_verlet_scheme(fplog, cr, inputrec, nstlist_cmdline, &mtop, box,
1091                           useGpuForNonbonded || (emulateGpuNonbonded == EmulateGpuNonbonded::Yes),
1092                           *hwinfo->cpuInfo);
1093
1094     const bool prefer1DAnd1PulseDD = (devFlags.enableGpuHaloExchange && useGpuForNonbonded);
1095     // This builder is necessary while we have multi-part construction
1096     // of DD. Before DD is constructed, we use the existence of
1097     // the builder object to indicate that further construction of DD
1098     // is needed.
1099     std::unique_ptr<DomainDecompositionBuilder> ddBuilder;
1100     if (useDomainDecomposition)
1101     {
1102         ddBuilder = std::make_unique<DomainDecompositionBuilder>(
1103                 mdlog, cr, domdecOptions, mdrunOptions, prefer1DAnd1PulseDD, mtop, *inputrec, box,
1104                 positionsFromStatePointer(globalState.get()));
1105     }
1106     else
1107     {
1108         /* PME, if used, is done on all nodes with 1D decomposition */
1109         cr->npmenodes = 0;
1110         cr->duty      = (DUTY_PP | DUTY_PME);
1111
1112         if (inputrec->ePBC == epbcSCREW)
1113         {
1114             gmx_fatal(FARGS, "pbc=screw is only implemented with domain decomposition");
1115         }
1116     }
1117
1118     // Produce the task assignment for this rank.
1119     GpuTaskAssignmentsBuilder gpuTaskAssignmentsBuilder;
1120     GpuTaskAssignments        gpuTaskAssignments = gpuTaskAssignmentsBuilder.build(
1121             gpuIdsToUse, userGpuTaskAssignment, *hwinfo, communicator, physicalNodeComm,
1122             nonbondedTarget, pmeTarget, bondedTarget, updateTarget, useGpuForNonbonded,
1123             useGpuForPme, thisRankHasDuty(cr, DUTY_PP),
1124             // TODO cr->duty & DUTY_PME should imply that a PME
1125             // algorithm is active, but currently does not.
1126             EEL_PME(inputrec->coulombtype) && thisRankHasDuty(cr, DUTY_PME));
1127
1128     // Get the device handles for the modules, nullptr when no task is assigned.
1129     gmx_device_info_t* nonbondedDeviceInfo = gpuTaskAssignments.initNonbondedDevice(cr);
1130     gmx_device_info_t* pmeDeviceInfo       = gpuTaskAssignments.initPmeDevice();
1131
1132     // TODO Initialize GPU streams here.
1133
1134     // TODO Currently this is always built, yet DD partition code
1135     // checks if it is built before using it. Probably it should
1136     // become an MDModule that is made only when another module
1137     // requires it (e.g. pull, CompEl, density fitting), so that we
1138     // don't update the local atom sets unilaterally every step.
1139     LocalAtomSetManager atomSets;
1140     if (ddBuilder)
1141     {
1142         // TODO Pass the GPU streams to ddBuilder to use in buffer
1143         // transfers (e.g. halo exchange)
1144         cr->dd = ddBuilder->build(&atomSets);
1145         // The builder's job is done, so destruct it
1146         ddBuilder.reset(nullptr);
1147         // Note that local state still does not exist yet.
1148     }
1149
1150     // The GPU update is decided here because we need to know whether the constraints or
1151     // SETTLEs can span accross the domain borders (i.e. whether or not update groups are
1152     // defined). This is only known after DD is initialized, hence decision on using GPU
1153     // update is done so late.
1154     try
1155     {
1156         const bool useUpdateGroups = cr->dd ? ddUsesUpdateGroups(*cr->dd) : false;
1157
1158         useGpuForUpdate = decideWhetherToUseGpuForUpdate(
1159                 useDomainDecomposition, useUpdateGroups, pmeRunMode, domdecOptions.numPmeRanks > 0,
1160                 useGpuForNonbonded, updateTarget, gpusWereDetected, *inputrec, mtop,
1161                 doEssentialDynamics, gmx_mtop_ftype_count(mtop, F_ORIRES) > 0,
1162                 replExParams.exchangeInterval > 0, doRerun, devFlags, mdlog);
1163     }
1164     GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1165
1166     const bool printHostName = (cr->nnodes > 1);
1167     gpuTaskAssignments.reportGpuUsage(mdlog, printHostName, useGpuForBonded, pmeRunMode, useGpuForUpdate);
1168
1169     // If the user chose a task assignment, give them some hints
1170     // where appropriate.
1171     if (!userGpuTaskAssignment.empty())
1172     {
1173         gpuTaskAssignments.logPerformanceHints(mdlog, ssize(gpuIdsToUse));
1174     }
1175
1176     if (PAR(cr))
1177     {
1178         /* After possible communicator splitting in make_dd_communicators.
1179          * we can set up the intra/inter node communication.
1180          */
1181         gmx_setup_nodecomm(fplog, cr);
1182     }
1183
1184 #if GMX_MPI
1185     if (isMultiSim(ms))
1186     {
1187         GMX_LOG(mdlog.warning)
1188                 .asParagraph()
1189                 .appendTextFormatted(
1190                         "This is simulation %d out of %d running as a composite GROMACS\n"
1191                         "multi-simulation job. Setup for this simulation:\n",
1192                         ms->sim, ms->nsim);
1193     }
1194     GMX_LOG(mdlog.warning)
1195             .appendTextFormatted("Using %d MPI %s\n", cr->nnodes,
1196 #    if GMX_THREAD_MPI
1197                                  cr->nnodes == 1 ? "thread" : "threads"
1198 #    else
1199                                  cr->nnodes == 1 ? "process" : "processes"
1200 #    endif
1201             );
1202     fflush(stderr);
1203 #endif
1204
1205     // If mdrun -pin auto honors any affinity setting that already
1206     // exists. If so, it is nice to provide feedback about whether
1207     // that existing affinity setting was from OpenMP or something
1208     // else, so we run this code both before and after we initialize
1209     // the OpenMP support.
1210     gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1211     /* Check and update the number of OpenMP threads requested */
1212     checkAndUpdateRequestedNumOpenmpThreads(&hw_opt, *hwinfo, cr, ms, physicalNodeComm.size_,
1213                                             pmeRunMode, mtop, *inputrec);
1214
1215     gmx_omp_nthreads_init(mdlog, cr, hwinfo->nthreads_hw_avail, physicalNodeComm.size_,
1216                           hw_opt.nthreads_omp, hw_opt.nthreads_omp_pme, !thisRankHasDuty(cr, DUTY_PP));
1217
1218     // Enable FP exception detection, but not in
1219     // Release mode and not for compilers with known buggy FP
1220     // exception support (clang with any optimization) or suspected
1221     // buggy FP exception support (gcc 7.* with optimization).
1222 #if !defined NDEBUG                                                                         \
1223         && !((defined __clang__ || (defined(__GNUC__) && !defined(__ICC) && __GNUC__ == 7)) \
1224              && defined __OPTIMIZE__)
1225     const bool bEnableFPE = true;
1226 #else
1227     const bool bEnableFPE = false;
1228 #endif
1229     // FIXME - reconcile with gmx_feenableexcept() call from CommandLineModuleManager::run()
1230     if (bEnableFPE)
1231     {
1232         gmx_feenableexcept();
1233     }
1234
1235     /* Now that we know the setup is consistent, check for efficiency */
1236     check_resource_division_efficiency(hwinfo, gpuTaskAssignments.thisRankHasAnyGpuTask(),
1237                                        mdrunOptions.ntompOptionIsSet, cr, mdlog);
1238
1239     /* getting number of PP/PME threads on this MPI / tMPI rank.
1240        PME: env variable should be read only on one node to make sure it is
1241        identical everywhere;
1242      */
1243     const int numThreadsOnThisRank = thisRankHasDuty(cr, DUTY_PP) ? gmx_omp_nthreads_get(emntNonbonded)
1244                                                                   : gmx_omp_nthreads_get(emntPME);
1245     checkHardwareOversubscription(numThreadsOnThisRank, cr->nodeid, *hwinfo->hardwareTopology,
1246                                   physicalNodeComm, mdlog);
1247
1248     // Enable Peer access between GPUs where available
1249     // Only for DD, only master PP rank needs to perform setup, and only if thread MPI plus
1250     // any of the GPU communication features are active.
1251     if (DOMAINDECOMP(cr) && MASTER(cr) && thisRankHasDuty(cr, DUTY_PP) && GMX_THREAD_MPI
1252         && (devFlags.enableGpuHaloExchange || devFlags.enableGpuPmePPComm))
1253     {
1254         setupGpuDevicePeerAccess(gpuIdsToUse, mdlog);
1255     }
1256
1257     if (hw_opt.threadAffinity != ThreadAffinity::Off)
1258     {
1259         /* Before setting affinity, check whether the affinity has changed
1260          * - which indicates that probably the OpenMP library has changed it
1261          * since we first checked).
1262          */
1263         gmx_check_thread_affinity_set(mdlog, &hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1264
1265         int numThreadsOnThisNode, intraNodeThreadOffset;
1266         analyzeThreadsOnThisNode(physicalNodeComm, numThreadsOnThisRank, &numThreadsOnThisNode,
1267                                  &intraNodeThreadOffset);
1268
1269         /* Set the CPU affinity */
1270         gmx_set_thread_affinity(mdlog, cr, &hw_opt, *hwinfo->hardwareTopology, numThreadsOnThisRank,
1271                                 numThreadsOnThisNode, intraNodeThreadOffset, nullptr);
1272     }
1273
1274     if (mdrunOptions.timingOptions.resetStep > -1)
1275     {
1276         GMX_LOG(mdlog.info)
1277                 .asParagraph()
1278                 .appendText(
1279                         "The -resetstep functionality is deprecated, and may be removed in a "
1280                         "future version.");
1281     }
1282     wcycle = wallcycle_init(fplog, mdrunOptions.timingOptions.resetStep, cr);
1283
1284     if (PAR(cr))
1285     {
1286         /* Master synchronizes its value of reset_counters with all nodes
1287          * including PME only nodes */
1288         int64_t reset_counters = wcycle_get_reset_counters(wcycle);
1289         gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1290         wcycle_set_reset_counters(wcycle, reset_counters);
1291     }
1292
1293     // Membrane embedding must be initialized before we call init_forcerec()
1294     if (doMembed)
1295     {
1296         if (MASTER(cr))
1297         {
1298             fprintf(stderr, "Initializing membed");
1299         }
1300         /* Note that membed cannot work in parallel because mtop is
1301          * changed here. Fix this if we ever want to make it run with
1302          * multiple ranks. */
1303         membed = init_membed(fplog, filenames.size(), filenames.data(), &mtop, inputrec,
1304                              globalState.get(), cr, &mdrunOptions.checkpointOptions.period);
1305     }
1306
1307     const bool                   thisRankHasPmeGpuTask = gpuTaskAssignments.thisRankHasPmeGpuTask();
1308     std::unique_ptr<MDAtoms>     mdAtoms;
1309     std::unique_ptr<gmx_vsite_t> vsite;
1310
1311     t_nrnb nrnb;
1312     if (thisRankHasDuty(cr, DUTY_PP))
1313     {
1314         mdModulesNotifier.notify(*cr);
1315         mdModulesNotifier.notify(&atomSets);
1316         mdModulesNotifier.notify(PeriodicBoundaryConditionType{ inputrec->ePBC });
1317         mdModulesNotifier.notify(SimulationTimeStep{ inputrec->delta_t });
1318         /* Initiate forcerecord */
1319         fr                 = new t_forcerec;
1320         fr->forceProviders = mdModules_->initForceProviders();
1321         init_forcerec(fplog, mdlog, fr, fcd, inputrec, &mtop, cr, box,
1322                       opt2fn("-table", filenames.size(), filenames.data()),
1323                       opt2fn("-tablep", filenames.size(), filenames.data()),
1324                       opt2fns("-tableb", filenames.size(), filenames.data()), *hwinfo,
1325                       nonbondedDeviceInfo, useGpuForBonded,
1326                       pmeRunMode == PmeRunMode::GPU && !thisRankHasDuty(cr, DUTY_PME), pforce, wcycle);
1327
1328         // TODO Move this to happen during domain decomposition setup,
1329         // once stream and event handling works well with that.
1330         // TODO remove need to pass local stream into GPU halo exchange - Redmine #3093
1331         if (havePPDomainDecomposition(cr) && prefer1DAnd1PulseDD && is1DAnd1PulseDD(*cr->dd))
1332         {
1333             GMX_RELEASE_ASSERT(devFlags.enableGpuBufferOps,
1334                                "Must use GMX_USE_GPU_BUFFER_OPS=1 to use GMX_GPU_DD_COMMS=1");
1335             void* streamLocal =
1336                     Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local);
1337             void* streamNonLocal =
1338                     Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal);
1339             GMX_LOG(mdlog.warning)
1340                     .asParagraph()
1341                     .appendTextFormatted(
1342                             "NOTE: This run uses the 'GPU halo exchange' feature, enabled by the "
1343                             "GMX_GPU_DD_COMMS environment variable.");
1344             cr->dd->gpuHaloExchange = std::make_unique<GpuHaloExchange>(
1345                     cr->dd, cr->mpi_comm_mysim, streamLocal, streamNonLocal);
1346         }
1347
1348         /* Initialize the mdAtoms structure.
1349          * mdAtoms is not filled with atom data,
1350          * as this can not be done now with domain decomposition.
1351          */
1352         mdAtoms = makeMDAtoms(fplog, mtop, *inputrec, thisRankHasPmeGpuTask);
1353         if (globalState && thisRankHasPmeGpuTask)
1354         {
1355             // The pinning of coordinates in the global state object works, because we only use
1356             // PME on GPU without DD or on a separate PME rank, and because the local state pointer
1357             // points to the global state object without DD.
1358             // FIXME: MD and EM separately set up the local state - this should happen in the same
1359             // function, which should also perform the pinning.
1360             changePinningPolicy(&globalState->x, pme_get_pinning_policy());
1361         }
1362
1363         /* Initialize the virtual site communication */
1364         vsite = initVsite(mtop, cr);
1365
1366         calc_shifts(box, fr->shift_vec);
1367
1368         /* With periodic molecules the charge groups should be whole at start up
1369          * and the virtual sites should not be far from their proper positions.
1370          */
1371         if (!inputrec->bContinuation && MASTER(cr) && !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1372         {
1373             /* Make molecules whole at start of run */
1374             if (fr->ePBC != epbcNONE)
1375             {
1376                 do_pbc_first_mtop(fplog, inputrec->ePBC, box, &mtop, globalState->x.rvec_array());
1377             }
1378             if (vsite)
1379             {
1380                 /* Correct initial vsite positions are required
1381                  * for the initial distribution in the domain decomposition
1382                  * and for the initial shell prediction.
1383                  */
1384                 constructVsitesGlobal(mtop, globalState->x);
1385             }
1386         }
1387
1388         if (EEL_PME(fr->ic->eeltype) || EVDW_PME(fr->ic->vdwtype))
1389         {
1390             ewaldcoeff_q  = fr->ic->ewaldcoeff_q;
1391             ewaldcoeff_lj = fr->ic->ewaldcoeff_lj;
1392         }
1393     }
1394     else
1395     {
1396         /* This is a PME only node */
1397
1398         GMX_ASSERT(globalState == nullptr,
1399                    "We don't need the state on a PME only rank and expect it to be unitialized");
1400
1401         ewaldcoeff_q  = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1402         ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1403     }
1404
1405     gmx_pme_t* sepPmeData = nullptr;
1406     // This reference hides the fact that PME data is owned by runner on PME-only ranks and by forcerec on other ranks
1407     GMX_ASSERT(thisRankHasDuty(cr, DUTY_PP) == (fr != nullptr),
1408                "Double-checking that only PME-only ranks have no forcerec");
1409     gmx_pme_t*& pmedata = fr ? fr->pmedata : sepPmeData;
1410
1411     // TODO should live in ewald module once its testing is improved
1412     //
1413     // Later, this program could contain kernels that might be later
1414     // re-used as auto-tuning progresses, or subsequent simulations
1415     // are invoked.
1416     PmeGpuProgramStorage pmeGpuProgram;
1417     if (thisRankHasPmeGpuTask)
1418     {
1419         pmeGpuProgram = buildPmeGpuProgram(pmeDeviceInfo);
1420     }
1421
1422     /* Initiate PME if necessary,
1423      * either on all nodes or on dedicated PME nodes only. */
1424     if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1425     {
1426         if (mdAtoms && mdAtoms->mdatoms())
1427         {
1428             nChargePerturbed = mdAtoms->mdatoms()->nChargePerturbed;
1429             if (EVDW_PME(inputrec->vdwtype))
1430             {
1431                 nTypePerturbed = mdAtoms->mdatoms()->nTypePerturbed;
1432             }
1433         }
1434         if (cr->npmenodes > 0)
1435         {
1436             /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1437             gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1438             gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1439         }
1440
1441         if (thisRankHasDuty(cr, DUTY_PME))
1442         {
1443             try
1444             {
1445                 pmedata = gmx_pme_init(cr, getNumPmeDomains(cr->dd), inputrec, nChargePerturbed != 0,
1446                                        nTypePerturbed != 0, mdrunOptions.reproducible, ewaldcoeff_q,
1447                                        ewaldcoeff_lj, gmx_omp_nthreads_get(emntPME), pmeRunMode,
1448                                        nullptr, pmeDeviceInfo, pmeGpuProgram.get(), mdlog);
1449             }
1450             GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
1451         }
1452     }
1453
1454
1455     if (EI_DYNAMICS(inputrec->eI))
1456     {
1457         /* Turn on signal handling on all nodes */
1458         /*
1459          * (A user signal from the PME nodes (if any)
1460          * is communicated to the PP nodes.
1461          */
1462         signal_handler_install();
1463     }
1464
1465     pull_t* pull_work = nullptr;
1466     if (thisRankHasDuty(cr, DUTY_PP))
1467     {
1468         /* Assumes uniform use of the number of OpenMP threads */
1469         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1470
1471         if (inputrec->bPull)
1472         {
1473             /* Initialize pull code */
1474             pull_work = init_pull(fplog, inputrec->pull, inputrec, &mtop, cr, &atomSets,
1475                                   inputrec->fepvals->init_lambda);
1476             if (inputrec->pull->bXOutAverage || inputrec->pull->bFOutAverage)
1477             {
1478                 initPullHistory(pull_work, &observablesHistory);
1479             }
1480             if (EI_DYNAMICS(inputrec->eI) && MASTER(cr))
1481             {
1482                 init_pull_output_files(pull_work, filenames.size(), filenames.data(), oenv, startingBehavior);
1483             }
1484         }
1485
1486         std::unique_ptr<EnforcedRotation> enforcedRotation;
1487         if (inputrec->bRot)
1488         {
1489             /* Initialize enforced rotation code */
1490             enforcedRotation =
1491                     init_rot(fplog, inputrec, filenames.size(), filenames.data(), cr, &atomSets,
1492                              globalState.get(), &mtop, oenv, mdrunOptions, startingBehavior);
1493         }
1494
1495         t_swap* swap = nullptr;
1496         if (inputrec->eSwapCoords != eswapNO)
1497         {
1498             /* Initialize ion swapping code */
1499             swap = init_swapcoords(fplog, inputrec,
1500                                    opt2fn_master("-swap", filenames.size(), filenames.data(), cr),
1501                                    &mtop, globalState.get(), &observablesHistory, cr, &atomSets,
1502                                    oenv, mdrunOptions, startingBehavior);
1503         }
1504
1505         /* Let makeConstraints know whether we have essential dynamics constraints. */
1506         auto constr = makeConstraints(mtop, *inputrec, pull_work, doEssentialDynamics, fplog,
1507                                       *mdAtoms->mdatoms(), cr, ms, &nrnb, wcycle, fr->bMolPBC);
1508
1509         /* Energy terms and groups */
1510         gmx_enerdata_t enerd(mtop.groups.groups[SimulationAtomGroupType::EnergyOutput].size(),
1511                              inputrec->fepvals->n_lambda);
1512
1513         /* Kinetic energy data */
1514         gmx_ekindata_t ekind;
1515         init_ekindata(fplog, &mtop, &(inputrec->opts), &ekind, inputrec->cos_accel);
1516
1517         /* Set up interactive MD (IMD) */
1518         auto imdSession =
1519                 makeImdSession(inputrec, cr, wcycle, &enerd, ms, &mtop, mdlog,
1520                                MASTER(cr) ? globalState->x.rvec_array() : nullptr, filenames.size(),
1521                                filenames.data(), oenv, mdrunOptions.imdOptions, startingBehavior);
1522
1523         if (DOMAINDECOMP(cr))
1524         {
1525             GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1526             /* This call is not included in init_domain_decomposition mainly
1527              * because fr->cginfo_mb is set later.
1528              */
1529             dd_init_bondeds(fplog, cr->dd, &mtop, vsite.get(), inputrec,
1530                             domdecOptions.checkBondedInteractions, fr->cginfo_mb);
1531         }
1532
1533         // TODO This is not the right place to manage the lifetime of
1534         // this data structure, but currently it's the easiest way to
1535         // make it work.
1536         MdrunScheduleWorkload runScheduleWork;
1537         // Also populates the simulation constant workload description.
1538         runScheduleWork.simulationWork = createSimulationWorkload(
1539                 useGpuForNonbonded, pmeRunMode, useGpuForBonded, useGpuForUpdate,
1540                 devFlags.enableGpuBufferOps, devFlags.enableGpuHaloExchange,
1541                 devFlags.enableGpuPmePPComm, haveEwaldSurfaceContribution(*inputrec));
1542
1543         std::unique_ptr<gmx::StatePropagatorDataGpu> stateGpu;
1544         if (gpusWereDetected
1545             && ((useGpuForPme && thisRankHasDuty(cr, DUTY_PME))
1546                 || runScheduleWork.simulationWork.useGpuBufferOps))
1547         {
1548             const void* pmeStream = pme_gpu_get_device_stream(fr->pmedata);
1549             const void* localStream =
1550                     fr->nbv->gpu_nbv != nullptr
1551                             ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::Local)
1552                             : nullptr;
1553             const void* nonLocalStream =
1554                     fr->nbv->gpu_nbv != nullptr
1555                             ? Nbnxm::gpu_get_command_stream(fr->nbv->gpu_nbv, InteractionLocality::NonLocal)
1556                             : nullptr;
1557             const void*        deviceContext = pme_gpu_get_device_context(fr->pmedata);
1558             const int          paddingSize   = pme_gpu_get_padding_size(fr->pmedata);
1559             GpuApiCallBehavior transferKind = (inputrec->eI == eiMD && !doRerun && !useModularSimulator)
1560                                                       ? GpuApiCallBehavior::Async
1561                                                       : GpuApiCallBehavior::Sync;
1562
1563             stateGpu = std::make_unique<gmx::StatePropagatorDataGpu>(
1564                     pmeStream, localStream, nonLocalStream, deviceContext, transferKind, paddingSize, wcycle);
1565             fr->stateGpu = stateGpu.get();
1566         }
1567
1568         GMX_ASSERT(stopHandlerBuilder_, "Runner must provide StopHandlerBuilder to simulator.");
1569         SimulatorBuilder simulatorBuilder;
1570
1571         // build and run simulator object based on user-input
1572         auto simulator = simulatorBuilder.build(
1573                 inputIsCompatibleWithModularSimulator, fplog, cr, ms, mdlog,
1574                 static_cast<int>(filenames.size()), filenames.data(), oenv, mdrunOptions,
1575                 startingBehavior, vsite.get(), constr.get(),
1576                 enforcedRotation ? enforcedRotation->getLegacyEnfrot() : nullptr, deform.get(),
1577                 mdModules_->outputProvider(), mdModules_->notifier(), inputrec, imdSession.get(),
1578                 pull_work, swap, &mtop, fcd, globalState.get(), &observablesHistory, mdAtoms.get(),
1579                 &nrnb, wcycle, fr, &enerd, &ekind, &runScheduleWork, replExParams, membed,
1580                 walltime_accounting, std::move(stopHandlerBuilder_), doRerun);
1581         simulator->run();
1582
1583         if (fr->pmePpCommGpu)
1584         {
1585             // destroy object since it is no longer required. (This needs to be done while the GPU context still exists.)
1586             fr->pmePpCommGpu.reset();
1587         }
1588
1589         if (inputrec->bPull)
1590         {
1591             finish_pull(pull_work);
1592         }
1593         finish_swapcoords(swap);
1594     }
1595     else
1596     {
1597         GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1598         /* do PME only */
1599         walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1600         gmx_pmeonly(pmedata, cr, &nrnb, wcycle, walltime_accounting, inputrec, pmeRunMode);
1601     }
1602
1603     wallcycle_stop(wcycle, ewcRUN);
1604
1605     /* Finish up, write some stuff
1606      * if rerunMD, don't write last frame again
1607      */
1608     finish_run(fplog, mdlog, cr, inputrec, &nrnb, wcycle, walltime_accounting,
1609                fr ? fr->nbv.get() : nullptr, pmedata, EI_DYNAMICS(inputrec->eI) && !isMultiSim(ms));
1610
1611     // clean up cycle counter
1612     wallcycle_destroy(wcycle);
1613
1614     // Free PME data
1615     if (pmedata)
1616     {
1617         gmx_pme_destroy(pmedata);
1618         pmedata = nullptr;
1619     }
1620
1621     // FIXME: this is only here to manually unpin mdAtoms->chargeA_ and state->x,
1622     // before we destroy the GPU context(s) in free_gpu_resources().
1623     // Pinned buffers are associated with contexts in CUDA.
1624     // As soon as we destroy GPU contexts after mdrunner() exits, these lines should go.
1625     mdAtoms.reset(nullptr);
1626     globalState.reset(nullptr);
1627     mdModules_.reset(nullptr); // destruct force providers here as they might also use the GPU
1628
1629     /* Free GPU memory and set a physical node tMPI barrier (which should eventually go away) */
1630     free_gpu_resources(fr, physicalNodeComm, hwinfo->gpu_info);
1631     free_gpu(nonbondedDeviceInfo);
1632     free_gpu(pmeDeviceInfo);
1633     done_forcerec(fr, mtop.molblock.size());
1634     sfree(fcd);
1635
1636     if (doMembed)
1637     {
1638         free_membed(membed);
1639     }
1640
1641     /* Does what it says */
1642     print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1643     walltime_accounting_destroy(walltime_accounting);
1644
1645     // Ensure log file content is written
1646     if (logFileHandle)
1647     {
1648         gmx_fio_flush(logFileHandle);
1649     }
1650
1651     /* Reset FPEs (important for unit tests) by disabling them. Assumes no
1652      * exceptions were enabled before function was called. */
1653     if (bEnableFPE)
1654     {
1655         gmx_fedisableexcept();
1656     }
1657
1658     auto rc = static_cast<int>(gmx_get_stop_condition());
1659
1660 #if GMX_THREAD_MPI
1661     /* we need to join all threads. The sub-threads join when they
1662        exit this function, but the master thread needs to be told to
1663        wait for that. */
1664     if (PAR(cr) && MASTER(cr))
1665     {
1666         tMPI_Finalize();
1667     }
1668 #endif
1669     return rc;
1670 }
1671
1672 Mdrunner::~Mdrunner()
1673 {
1674     // Clean up of the Manager.
1675     // This will end up getting called on every thread-MPI rank, which is unnecessary,
1676     // but okay as long as threads synchronize some time before adding or accessing
1677     // a new set of restraints.
1678     if (restraintManager_)
1679     {
1680         restraintManager_->clear();
1681         GMX_ASSERT(restraintManager_->countRestraints() == 0,
1682                    "restraints added during runner life time should be cleared at runner "
1683                    "destruction.");
1684     }
1685 };
1686
1687 void Mdrunner::addPotential(std::shared_ptr<gmx::IRestraintPotential> puller, const std::string& name)
1688 {
1689     GMX_ASSERT(restraintManager_, "Mdrunner must have a restraint manager.");
1690     // Not sure if this should be logged through the md logger or something else,
1691     // but it is helpful to have some sort of INFO level message sent somewhere.
1692     //    std::cout << "Registering restraint named " << name << std::endl;
1693
1694     // When multiple restraints are used, it may be wasteful to register them separately.
1695     // Maybe instead register an entire Restraint Manager as a force provider.
1696     restraintManager_->addToSpec(std::move(puller), name);
1697 }
1698
1699 Mdrunner::Mdrunner(std::unique_ptr<MDModules> mdModules) : mdModules_(std::move(mdModules)) {}
1700
1701 Mdrunner::Mdrunner(Mdrunner&&) noexcept = default;
1702
1703 //NOLINTNEXTLINE(performance-noexcept-move-constructor) working around GCC bug 58265
1704 Mdrunner& Mdrunner::operator=(Mdrunner&& /*handle*/) noexcept(BUGFREE_NOEXCEPT_STRING) = default;
1705
1706 class Mdrunner::BuilderImplementation
1707 {
1708 public:
1709     BuilderImplementation() = delete;
1710     BuilderImplementation(std::unique_ptr<MDModules> mdModules, compat::not_null<SimulationContext*> context);
1711     ~BuilderImplementation();
1712
1713     BuilderImplementation& setExtraMdrunOptions(const MdrunOptions& options,
1714                                                 real                forceWarningThreshold,
1715                                                 StartingBehavior    startingBehavior);
1716
1717     void addDomdec(const DomdecOptions& options);
1718
1719     void addVerletList(int nstlist);
1720
1721     void addReplicaExchange(const ReplicaExchangeParameters& params);
1722
1723     void addNonBonded(const char* nbpu_opt);
1724
1725     void addPME(const char* pme_opt_, const char* pme_fft_opt_);
1726
1727     void addBondedTaskAssignment(const char* bonded_opt);
1728
1729     void addUpdateTaskAssignment(const char* update_opt);
1730
1731     void addHardwareOptions(const gmx_hw_opt_t& hardwareOptions);
1732
1733     void addFilenames(ArrayRef<const t_filenm> filenames);
1734
1735     void addOutputEnvironment(gmx_output_env_t* outputEnvironment);
1736
1737     void addLogFile(t_fileio* logFileHandle);
1738
1739     void addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder);
1740
1741     Mdrunner build();
1742
1743 private:
1744     // Default parameters copied from runner.h
1745     // \todo Clarify source(s) of default parameters.
1746
1747     const char* nbpu_opt_    = nullptr;
1748     const char* pme_opt_     = nullptr;
1749     const char* pme_fft_opt_ = nullptr;
1750     const char* bonded_opt_  = nullptr;
1751     const char* update_opt_  = nullptr;
1752
1753     MdrunOptions mdrunOptions_;
1754
1755     DomdecOptions domdecOptions_;
1756
1757     ReplicaExchangeParameters replicaExchangeParameters_;
1758
1759     //! Command-line override for the duration of a neighbor list with the Verlet scheme.
1760     int nstlist_ = 0;
1761
1762     //! Multisim communicator handle.
1763     gmx_multisim_t* multiSimulation_;
1764
1765     //! mdrun communicator
1766     MPI_Comm communicator_ = MPI_COMM_NULL;
1767
1768     //! Print a warning if any force is larger than this (in kJ/mol nm).
1769     real forceWarningThreshold_ = -1;
1770
1771     //! Whether the simulation will start afresh, or restart with/without appending.
1772     StartingBehavior startingBehavior_ = StartingBehavior::NewSimulation;
1773
1774     //! The modules that comprise the functionality of mdrun.
1775     std::unique_ptr<MDModules> mdModules_;
1776
1777     //! \brief Parallelism information.
1778     gmx_hw_opt_t hardwareOptions_;
1779
1780     //! filename options for simulation.
1781     ArrayRef<const t_filenm> filenames_;
1782
1783     /*! \brief Handle to output environment.
1784      *
1785      * \todo gmx_output_env_t needs lifetime management.
1786      */
1787     gmx_output_env_t* outputEnvironment_ = nullptr;
1788
1789     /*! \brief Non-owning handle to MD log file.
1790      *
1791      * \todo Context should own output facilities for client.
1792      * \todo Improve log file handle management.
1793      * \internal
1794      * Code managing the FILE* relies on the ability to set it to
1795      * nullptr to check whether the filehandle is valid.
1796      */
1797     t_fileio* logFileHandle_ = nullptr;
1798
1799     /*!
1800      * \brief Builder for simulation stop signal handler.
1801      */
1802     std::unique_ptr<StopHandlerBuilder> stopHandlerBuilder_ = nullptr;
1803 };
1804
1805 Mdrunner::BuilderImplementation::BuilderImplementation(std::unique_ptr<MDModules> mdModules,
1806                                                        compat::not_null<SimulationContext*> context) :
1807     mdModules_(std::move(mdModules))
1808 {
1809     communicator_    = context->communicator_;
1810     multiSimulation_ = context->multiSimulation_.get();
1811 }
1812
1813 Mdrunner::BuilderImplementation::~BuilderImplementation() = default;
1814
1815 Mdrunner::BuilderImplementation&
1816 Mdrunner::BuilderImplementation::setExtraMdrunOptions(const MdrunOptions&    options,
1817                                                       const real             forceWarningThreshold,
1818                                                       const StartingBehavior startingBehavior)
1819 {
1820     mdrunOptions_          = options;
1821     forceWarningThreshold_ = forceWarningThreshold;
1822     startingBehavior_      = startingBehavior;
1823     return *this;
1824 }
1825
1826 void Mdrunner::BuilderImplementation::addDomdec(const DomdecOptions& options)
1827 {
1828     domdecOptions_ = options;
1829 }
1830
1831 void Mdrunner::BuilderImplementation::addVerletList(int nstlist)
1832 {
1833     nstlist_ = nstlist;
1834 }
1835
1836 void Mdrunner::BuilderImplementation::addReplicaExchange(const ReplicaExchangeParameters& params)
1837 {
1838     replicaExchangeParameters_ = params;
1839 }
1840
1841 Mdrunner Mdrunner::BuilderImplementation::build()
1842 {
1843     auto newRunner = Mdrunner(std::move(mdModules_));
1844
1845     newRunner.mdrunOptions     = mdrunOptions_;
1846     newRunner.pforce           = forceWarningThreshold_;
1847     newRunner.startingBehavior = startingBehavior_;
1848     newRunner.domdecOptions    = domdecOptions_;
1849
1850     // \todo determine an invariant to check or confirm that all gmx_hw_opt_t objects are valid
1851     newRunner.hw_opt = hardwareOptions_;
1852
1853     // No invariant to check. This parameter exists to optionally override other behavior.
1854     newRunner.nstlist_cmdline = nstlist_;
1855
1856     newRunner.replExParams = replicaExchangeParameters_;
1857
1858     newRunner.filenames = filenames_;
1859
1860     newRunner.communicator = communicator_;
1861
1862     // nullptr is a valid value for the multisim handle
1863     newRunner.ms = multiSimulation_;
1864
1865     // \todo Clarify ownership and lifetime management for gmx_output_env_t
1866     // \todo Update sanity checking when output environment has clearly specified invariants.
1867     // Initialization and default values for oenv are not well specified in the current version.
1868     if (outputEnvironment_)
1869     {
1870         newRunner.oenv = outputEnvironment_;
1871     }
1872     else
1873     {
1874         GMX_THROW(gmx::APIError(
1875                 "MdrunnerBuilder::addOutputEnvironment() is required before build()"));
1876     }
1877
1878     newRunner.logFileHandle = logFileHandle_;
1879
1880     if (nbpu_opt_)
1881     {
1882         newRunner.nbpu_opt = nbpu_opt_;
1883     }
1884     else
1885     {
1886         GMX_THROW(gmx::APIError("MdrunnerBuilder::addNonBonded() is required before build()"));
1887     }
1888
1889     if (pme_opt_ && pme_fft_opt_)
1890     {
1891         newRunner.pme_opt     = pme_opt_;
1892         newRunner.pme_fft_opt = pme_fft_opt_;
1893     }
1894     else
1895     {
1896         GMX_THROW(gmx::APIError("MdrunnerBuilder::addElectrostatics() is required before build()"));
1897     }
1898
1899     if (bonded_opt_)
1900     {
1901         newRunner.bonded_opt = bonded_opt_;
1902     }
1903     else
1904     {
1905         GMX_THROW(gmx::APIError(
1906                 "MdrunnerBuilder::addBondedTaskAssignment() is required before build()"));
1907     }
1908
1909     if (update_opt_)
1910     {
1911         newRunner.update_opt = update_opt_;
1912     }
1913     else
1914     {
1915         GMX_THROW(gmx::APIError(
1916                 "MdrunnerBuilder::addUpdateTaskAssignment() is required before build()  "));
1917     }
1918
1919
1920     newRunner.restraintManager_ = std::make_unique<gmx::RestraintManager>();
1921
1922     if (stopHandlerBuilder_)
1923     {
1924         newRunner.stopHandlerBuilder_ = std::move(stopHandlerBuilder_);
1925     }
1926     else
1927     {
1928         newRunner.stopHandlerBuilder_ = std::make_unique<StopHandlerBuilder>();
1929     }
1930
1931     return newRunner;
1932 }
1933
1934 void Mdrunner::BuilderImplementation::addNonBonded(const char* nbpu_opt)
1935 {
1936     nbpu_opt_ = nbpu_opt;
1937 }
1938
1939 void Mdrunner::BuilderImplementation::addPME(const char* pme_opt, const char* pme_fft_opt)
1940 {
1941     pme_opt_     = pme_opt;
1942     pme_fft_opt_ = pme_fft_opt;
1943 }
1944
1945 void Mdrunner::BuilderImplementation::addBondedTaskAssignment(const char* bonded_opt)
1946 {
1947     bonded_opt_ = bonded_opt;
1948 }
1949
1950 void Mdrunner::BuilderImplementation::addUpdateTaskAssignment(const char* update_opt)
1951 {
1952     update_opt_ = update_opt;
1953 }
1954
1955 void Mdrunner::BuilderImplementation::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
1956 {
1957     hardwareOptions_ = hardwareOptions;
1958 }
1959
1960 void Mdrunner::BuilderImplementation::addFilenames(ArrayRef<const t_filenm> filenames)
1961 {
1962     filenames_ = filenames;
1963 }
1964
1965 void Mdrunner::BuilderImplementation::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
1966 {
1967     outputEnvironment_ = outputEnvironment;
1968 }
1969
1970 void Mdrunner::BuilderImplementation::addLogFile(t_fileio* logFileHandle)
1971 {
1972     logFileHandle_ = logFileHandle;
1973 }
1974
1975 void Mdrunner::BuilderImplementation::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
1976 {
1977     stopHandlerBuilder_ = std::move(builder);
1978 }
1979
1980 MdrunnerBuilder::MdrunnerBuilder(std::unique_ptr<MDModules>           mdModules,
1981                                  compat::not_null<SimulationContext*> context) :
1982     impl_{ std::make_unique<Mdrunner::BuilderImplementation>(std::move(mdModules), context) }
1983 {
1984 }
1985
1986 MdrunnerBuilder::~MdrunnerBuilder() = default;
1987
1988 MdrunnerBuilder& MdrunnerBuilder::addSimulationMethod(const MdrunOptions&    options,
1989                                                       real                   forceWarningThreshold,
1990                                                       const StartingBehavior startingBehavior)
1991 {
1992     impl_->setExtraMdrunOptions(options, forceWarningThreshold, startingBehavior);
1993     return *this;
1994 }
1995
1996 MdrunnerBuilder& MdrunnerBuilder::addDomainDecomposition(const DomdecOptions& options)
1997 {
1998     impl_->addDomdec(options);
1999     return *this;
2000 }
2001
2002 MdrunnerBuilder& MdrunnerBuilder::addNeighborList(int nstlist)
2003 {
2004     impl_->addVerletList(nstlist);
2005     return *this;
2006 }
2007
2008 MdrunnerBuilder& MdrunnerBuilder::addReplicaExchange(const ReplicaExchangeParameters& params)
2009 {
2010     impl_->addReplicaExchange(params);
2011     return *this;
2012 }
2013
2014 MdrunnerBuilder& MdrunnerBuilder::addNonBonded(const char* nbpu_opt)
2015 {
2016     impl_->addNonBonded(nbpu_opt);
2017     return *this;
2018 }
2019
2020 MdrunnerBuilder& MdrunnerBuilder::addElectrostatics(const char* pme_opt, const char* pme_fft_opt)
2021 {
2022     // The builder method may become more general in the future, but in this version,
2023     // parameters for PME electrostatics are both required and the only parameters
2024     // available.
2025     if (pme_opt && pme_fft_opt)
2026     {
2027         impl_->addPME(pme_opt, pme_fft_opt);
2028     }
2029     else
2030     {
2031         GMX_THROW(
2032                 gmx::InvalidInputError("addElectrostatics() arguments must be non-null pointers."));
2033     }
2034     return *this;
2035 }
2036
2037 MdrunnerBuilder& MdrunnerBuilder::addBondedTaskAssignment(const char* bonded_opt)
2038 {
2039     impl_->addBondedTaskAssignment(bonded_opt);
2040     return *this;
2041 }
2042
2043 MdrunnerBuilder& MdrunnerBuilder::addUpdateTaskAssignment(const char* update_opt)
2044 {
2045     impl_->addUpdateTaskAssignment(update_opt);
2046     return *this;
2047 }
2048
2049 Mdrunner MdrunnerBuilder::build()
2050 {
2051     return impl_->build();
2052 }
2053
2054 MdrunnerBuilder& MdrunnerBuilder::addHardwareOptions(const gmx_hw_opt_t& hardwareOptions)
2055 {
2056     impl_->addHardwareOptions(hardwareOptions);
2057     return *this;
2058 }
2059
2060 MdrunnerBuilder& MdrunnerBuilder::addFilenames(ArrayRef<const t_filenm> filenames)
2061 {
2062     impl_->addFilenames(filenames);
2063     return *this;
2064 }
2065
2066 MdrunnerBuilder& MdrunnerBuilder::addOutputEnvironment(gmx_output_env_t* outputEnvironment)
2067 {
2068     impl_->addOutputEnvironment(outputEnvironment);
2069     return *this;
2070 }
2071
2072 MdrunnerBuilder& MdrunnerBuilder::addLogFile(t_fileio* logFileHandle)
2073 {
2074     impl_->addLogFile(logFileHandle);
2075     return *this;
2076 }
2077
2078 MdrunnerBuilder& MdrunnerBuilder::addStopHandlerBuilder(std::unique_ptr<StopHandlerBuilder> builder)
2079 {
2080     impl_->addStopHandlerBuilder(std::move(builder));
2081     return *this;
2082 }
2083
2084 MdrunnerBuilder::MdrunnerBuilder(MdrunnerBuilder&&) noexcept = default;
2085
2086 MdrunnerBuilder& MdrunnerBuilder::operator=(MdrunnerBuilder&&) noexcept = default;
2087
2088 } // namespace gmx